My Community

Hello!!!

I'm trying to understand how machine learning force field works in vasp.

I thought the energy, forces, stress tensor and their uncertainties are computed by bayseian inteference during MD simulation.
However, after I fitted Force field, i saw the energy discrepancy is always constant between first principles calculation and force field energy.
I was wondering if the training procedure give more weights to forces, rather than energies or other properties? Then, it makes sense that the energy is always off by a constant value.

Thank you in advance!!!

Hello,

I didn't quite understand what energies you are trying to compare. Is it ab-initio MD simulation vs. FF MD simulation? Could you attach the relevant files according to the forum guidelines?

Hello Alexey,

Thank you for your response and asking for clarifications!

I wanted to compare single point energies beetween DFT and trained force field for different configurations. I have observed that the discrepancy between these two energies(DFT vs trained MLFF) are almost always similar in the fact that MLFF energies are always underestimated compared to DFT energies by roughly the same difference.
I wanted to understand why this trend is happening.

I assume this is because in the FF training code, the DFT is called only above threshold of forces only.
I just wanted to know if the discrepancy results from the threshold of forces. I have attached the relvant files I have used.

Thank you for providing the input files. Could you also attach the output files from one of such comparisons (DFT and MLFF)?

Thanks Alexey!
Attachment is an example preliminary result I have organized. (I also included the input files I have used in VASP, also with the relevant output file).

Please let me know if you have any idea or suggestion regarding my question earlier! Thanks a lot, alexey.

Thank you for providing the files.

However, after I fitted Force field, i saw the energy discrepancy is always constant between first principles calculation and force field energy.

From the plots you provided I don't see that the energy difference is always constant. Indeed, the FF energies are lower than the ab-initio one. You can try to refit your FF with ML_MODE=reft and increase the weight of the potential energy ML_WTOTEN=30. This should bring the FF energy closer to the ab-initio one.

Hi Alexey!

Thanks a lot for continuous answers and advice!
I didn't know the tag ML_WTOTEN so far.

Thank you, again!!